Incremental Adversarial Learning for Optimal Path Planning (1809.09433v1)

Abstract: Path planning plays an essential role in many areas of robotics. Various planning techniques have been presented, either focusing on learning a specific task from demonstrations or retrieving trajectories by optimizing for hand-crafted cost functions which are well defined a priori. In this work, we present an incremental adversarial learning-based framework that allows inferring implicit behaviour, i.e. the natural characteristic of a set of given trajectories. To achieve adversarial learning, a zero-sum game is constructed between a planning algorithm and an adversary - the discriminator. We employ the discriminator within an optimal motion planning algorithm, such that costs can be learned and optimized iteratively, improving the integration of implicit behavior. By combining a cost-based planning approach with trained intrinsic behaviour, this can be be integrated also with other constraints such as obstacles or general cost factors within a single planning framework. We demonstrate the proposed method on a dataset for collision avoidance, as well as for the generation of human-like trajectories from motion capture data. Our results show that incremental adversarial learning is able to generate paths that reflect the natural implicit behaviour of a dataset, with the ability to improve on performance using iterative learning and generation.